[cmake-developers] Modifying RPATH feature to run tests uninstalled

[Stephen Kelly]

Hi there,

One of the macros we have in KDE creates shell scripts to initialize 
environment variables needed on various platforms so that tests can be run 
before installation. If RPATH is enabled, the scripts are not needed, but 
RPATH is sometimes disabled, so the scripts are the solution to that.

With the modularization of kdelibs into KDE frameworks (multiple independent 
pieces), we wondered what to do with that macro and how we can retain its 
features easily after making kdelibs non-monolithic.

I chatted with debian packager Sune Vuorela this evening about this. The 
linked thread is long and contains long emails, so all of the context is in 
the IRC log.

<steveire> svuorela: ping?
<steveire> I'm wondering about debian packages.
<steveire> Is it normal to run tests before installing them?
<steveire> Grep for 'Sune' here: 
http://thread.gmane.org/gmane.comp.kde.devel.buildsystem/6961/focus=201
<svuorela> steveire: it depends on the package. we aren't running kdelibs 
tests, but we are running libc tests
<steveire> svuorela: And do you run them before installation?
<svuorela> steveire: I think so yes. but even if run after installation it 
is in a weird prefix (DESTDIR)
<steveire> By coincidence I got an email yesterday from bricks (who packages 
Grantlee) and he was asking about running tests before installation. He hit 
the issue in that thread for the exact reason that is there (ie debian 
disabled RPATH). Is that the 'normal' way to run tests in debian packaging 
systems?
<svuorela> steveire: so even if RPATH was set you coulddn't easily actually 
find the installed bits ...
<steveire> ... if the install location isn't already in LD_LIBRARY_PATH?
<svuorela> ...but as a normal user, I would expect that make intsall is 
something I run after a succsessfull make test
<steveire> svuorela: What do you think of Alex' suggestion? Or should 
CMAKE_SKIP_RPATH only disable the RPATH for the install tree ?
<steveire> Debian packagers still wouldn't allow the RPATH to be used in the 
binary dir, would they?
<steveire> So it still wouldn't be possbile to run them uninstalled?
<svuorela> at least from a debian pov, it is the installed version of things 
that matters. if  there is a intermediate RPATH at a stage during 
compilation no one would care.
<svuorela> and I actually think that it makes quite good sense to have RPATH 
on during compilation
<steveire> Right, which is why kde creates wrapper scripts to run unit tests 
with env vars set to find the libraries.
<steveire> It looks like this is the most likely solution: http://www.mail-
archive.com/cmake at cmake.org/msg12056.html What do you think?
<steveire> We could add LD_LIBRARY_PATH/DYLD_LIBRARY_PATH/PATH in a command 
like that.
<svuorela> steveire: I think it makes good sense both to have RPATHs in the 
build tree (and strip them on install if -DCMAKE_SKIP_RPATH) and to have a 
way of setting env vars for make test
<steveire> svuorela: Ok, do you mind if I post some of this discussion to 
the cmake-developers list?
<svuorela> steveire: feel free.


So the suggestion is

a) Change CMAKE_SKIP_RPATH to only skip RPATH when installing, but not when 
building (I don't know if that's easy in CMake), or alternatively provide 
another CMAKE_ variable to do that.

b) Implement something like:
set_property(TARGET foo APPEND PROPERTY TEST_ENVIRONMENT "foo=bar")


Possibly both a) and b).

Thoughts/comments?

Thanks,

Steve.